1. OpenBMB Releases MiniCPM5-2B On-Device Model
OpenBMB's MiniCPM5-2B is a highly optimized dense causal language model designed to run efficiently on-device. Built using a training process that incorporated 400 billion tokens of supervised fine-tuning (SFT) and on-policy distillation of 16 reinforcement learning experts, the model achieves strong performance in tool use, coding agents, and long-context retrieval. It outperforms the Qwen3.5-4B baseline on average across 34 benchmarks and demonstrates high token efficiency, using only 19k output tokens per task on the Artificial Analysis Intelligence Index.
- • MiniCPM5-2B is a 2.52 billion parameter dense model released under the Apache 2.0 license.
- • The model features a native context window of 131,072 tokens and a 42-layer architecture with grouped-query attention.
- • It scored 15 on the Artificial Analysis Intelligence Index v4.2, the highest score for any open-weights model under 4 billion parameters.
- • The model is compatible with standard runtimes including vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, and MLX.
- • OpenBMB released the model weights alongside intermediate checkpoints and datasets like Ultra-FineWeb and UltraX.
This model provides developers with a highly capable, token-efficient, and fully open-weights option for local agentic workflows and long-context retrieval on consumer hardware.
2. Tencent Releases EVIE Visual Document Retrieval Models
Tencent's new EVIE-8B and EVIE-4.5B models target the challenges of visual document retrieval in RAG pipelines. By representing document pages visually, they preserve critical layout, tabular, and typographic information that text-only parsers often lose. The EVIE-4.5B model is particularly optimized for production, featuring Prefix-MRL elasticity for dynamic dimension truncation and advanced clustering that compresses pages down to 32 vectors, drastically reducing vector database storage costs.
- • EVIE-8B achieves a state-of-the-art retrieval performance of 66.75 nDCG@10 on the ViDoRe V3 benchmark.
- • The models utilize 4096-dimensional representations to preserve fine-grained details like typography, charts, and tables.
- • EVIE-4.5B features Prefix-MRL elasticity, allowing runtime truncation of dimensions between 64 and 2048.
- • EVIE-4.5B uses Hierarchical Agglomerative Clustering to compress token counts to 32 vectors per page.
- • The compression reduces index storage requirements to 3.81 GiB per million pages.
These models enable developers to build highly accurate visual RAG pipelines that preserve complex document layouts, charts, and tables without massive storage overhead.
3. Engrim Launches Local-First SQLite Memory Engine for AI CLIs
Created by Tim Gordon, Engrim addresses the challenge of context fragmentation when using multiple AI coding assistants on the same codebase. By acting as a universal, local-first memory layer, it ensures that architectural decisions and project states are preserved across tools like Cursor and Claude Code. The engine combines keyword search with lightweight vector embeddings for fast context loading, and its offline-first architecture ensures that sensitive codebase metadata remains secure on the developer's local machine.
- • Engrim is a project-scoped, local-first memory engine licensed under the MIT license.
- • It supports multiple environments including Claude Code, Cursor, Windsurf, and Google Antigravity.
- • The engine uses a hybrid search combining SQLite FTS5 keyword search and model2vec static vector embeddings.
- • It operates entirely offline, storing data in a restricted-permission SQLite file at ~/.engrim/memory.db.
- • Engrim tracks the provenance of each memory entry to identify which agent or user created it.
Developers can use Engrim to maintain consistent context and state across different coding tools like Claude Code, Cursor, and Windsurf without leaking data offline.
4. LLM-as-a-Verifier Framework Released for Agent Feedback
The open-source LLM-as-a-Verifier framework offers a training-free method to enhance the reliability of AI agents. By generating detailed, step-by-step feedback, the framework allows developers to implement test-time scaling and progress tracking directly within their agentic pipelines. The system has demonstrated state-of-the-art results across diverse domains, including software engineering and robotics, making it a versatile tool for developers building complex multi-step agents.
- • LLM-as-a-Verifier is a general-purpose, training-free framework available on GitHub.
- • The framework provides fine-grained feedback to guide and correct agentic workflows.
- • It achieves state-of-the-art performance across coding, robotics, and medical agent benchmarks.
- • Feedback from the verifier can be applied directly to test-time scaling and reinforcement learning.
Developers can integrate this framework to implement test-time scaling, progress tracking, and reinforcement learning for their agents without the overhead of fine-tuning.
5. FreeCAD MCP Server Integrates AI Assistants with CAD Workflows
The freecad-mcp project brings the Model Context Protocol to computer-aided design, enabling developers to control FreeCAD via local AI agents. By setting up the MCP server and starting the companion RPC server within FreeCAD, developers can pass natural language prompts to local models to execute complex geometric design tasks. Furthermore, by utilizing vision-capable GGUF models, the assistant can visually inspect FreeCAD screenshots to verify that its geometric operations were executed correctly.
- • The freecad-mcp project is an addon that integrates FreeCAD with AI modeling assistants.
- • It supports using local runtimes like llama.cpp or pi as modeling assistants via an MCP server configuration.
- • The system includes an RPC server within FreeCAD to handle communication with the AI assistant.
- • Loading a multimedia-capable model with the --mmproj option allows the AI to analyze screenshots and verify geometric operations.
- • Prerequisites include FreeCAD, llama.cpp, a GGUF model, and the uv package manager.
This integration allows developers to build agentic workflows that can generate, modify, and verify physical 3D models using natural language commands.